Two months later, the A-list actress Scarlett Johansson sent a legal letter to OpenAI when she realized its new ChatGPT voice was "eerily similar" to hers. This story is from the WIRED World in 2025, our annual trends briefing. People understand that this technology is being built on their data, often without our permission. It's no wonder that public confidence in AI is declining. A recent study by Pew Research shows that more than half of Americans are more concerned than they are excited about AI, a sentiment echoed by a majority of people from Central and South American, African, and Middle Eastern countries in a World Risk Poll.

Alignment of pretrained LLMs using instruction-based datasets is critical for creating fine-tuned models that reflect human preference. A growing number of alignment-based fine-tuning algorithms and benchmarks emerged recently, fueling the efforts on effective alignments of pre-trained LLMs to ensure helpful, harmless, and honest answers from both open-source and closed-source LLMs. This paper tackles this problem by developing an alignment fusion approach, coined as $H^3$Fusion, with three unique characteristics. First, $H^3$Fusion ensembles multiple individually aligned LLMs to create a final fine-tuned alignment model with enhanced capabilities beyond those of individual models, delivering robust alignment through promoting helpful, harmless, honest fusion. Second, $H^3$Fusion leverages the mixture-of-experts (MoE) methodology in two steps. We first freeze the multi-head attention weights of each individual model while tuning the FFN layer during alignment fusion. Then we merge the aligned model weights with an expert router according to the type of input instruction and dynamically select a subset of experts that are best suited for producing the output response. Finally, we boost the performance of the resulting $H^3$3Fusion model by introducing gating loss and regularization terms. The former penalizes the selection errors of the expert-router, and the latter mediates the expert weights drifting during fine-tuning and dynamically adjusts the fusion behavior of the resulting model by canalizing the activations on the experts. Extensive evaluations on three benchmark datasets show that $H^3$3Fusion is more helpful, less harmful, and more honest from two aspects: it outperforms each individually aligned model by $11.37\%$, and it provides stronger robustness compared to the state-of-the-art LLM ensemble approaches by $13.77\%$. Code is available at github.com/sftekin/h3fusion.

AI evaluations are an important component of the AI governance toolkit, underlying current approaches to safety cases for preventing catastrophic risks. Our paper examines what these evaluations can and cannot tell us. Evaluations can establish lower bounds on AI capabilities and assess certain misuse risks given sufficient effort from evaluators. Unfortunately, evaluations face fundamental limitations that cannot be overcome within the current paradigm. These include an inability to establish upper bounds on capabilities, reliably forecast future model capabilities, or robustly assess risks from autonomous AI systems. This means that while evaluations are valuable tools, we should not rely on them as our main way of ensuring AI systems are safe. We conclude with recommendations for incremental improvements to frontier AI safety, while acknowledging these fundamental limitations remain unsolved.

Large language models (LLMs) have demonstrated remarkable capabilities in complex reasoning and text generation. However, these models can inadvertently generate unsafe or biased responses when prompted with problematic inputs, raising significant ethical and practical concerns for real-world deployment. This research addresses the critical challenge of developing language models that generate both helpful and harmless content, navigating the delicate balance between model performance and safety. We demonstrate that incorporating safety-related instructions during the instruction-tuning of pre-trained models significantly reduces toxic responses to unsafe prompts without compromising performance on helpfulness datasets. We found Direct Preference Optimization (DPO) to be particularly effective, outperforming both SIT and RAFT by leveraging both chosen and rejected responses for learning. Our approach increased safe responses from 40$\%$ to over 90$\%$ across various harmfulness benchmarks. In addition, we discuss a rigorous evaluation framework encompassing specialized metrics and diverse datasets for safety and helpfulness tasks ensuring a comprehensive assessment of the model's capabilities.

Transformer-based large-scale pre-trained models achieve great success, and fine-tuning, which tunes a pre-trained model on a task-specific dataset, is the standard practice to utilize these models for downstream tasks. Recent work has developed adapter-tuning, but these approaches either still require a relatively high resource usage. Through our investigation, we show that each adapter in adapter-tuning does not have the same impact on task performance and resource usage. Based on our findings, we propose SAFE, which gradually freezes less-important adapters that do not contribute to adaptation during the early training steps. In our experiments, SAFE reduces memory usage, computation amount, and training time by 42.85\%, 34.59\%, and 11.82\%, respectively, while achieving comparable or better performance compared to the baseline. We also demonstrate that SAFE induces regularization effect, thereby smoothing the loss landscape.

This chapter explores the influence of Artificial Intelligence (AI) on digital democracy, focusing on four main areas: citizenship, participation, representation, and the public sphere. It traces the evolution from electronic to virtual and network democracy, underscoring how each stage has broadened democratic engagement through technology. Focusing on digital citizenship, the chapter examines how AI can improve online engagement and promote ethical behaviour while posing privacy risks and fostering identity stereotyping. Regarding political participation, it highlights AI's dual role in mobilising civic actions and spreading misinformation. Regarding representation, AI's involvement in electoral processes can enhance voter registration, e-voting, and the efficiency of result tabulation but raises concerns regarding privacy and public trust. Also, AI's predictive capabilities shift the dynamics of political competition, posing ethical questions about manipulation and the legitimacy of democracy. Finally, the chapter examines how integrating AI and digital technologies can facilitate democratic political advocacy and personalised communication. However, this also comes with higher risks of misinformation and targeted propaganda.

Artificial Intelligence has proven to be a transformative tool for advancing scientific research across a wide range of disciplines. However, a significant gap still exists between AI and scientific communities, limiting the full potential of AI methods in driving broad scientific discovery. Existing efforts in bridging this gap have often relied on qualitative examination of small samples of literature, offering a limited perspective on the broader AI4Science landscape. In this work, we present a large-scale analysis of the AI4Science literature, starting by using large language models to identify scientific problems and AI methods in publications from top science and AI venues. Leveraging this new dataset, we quantitatively highlight key disparities between AI methods and scientific problems in this integrated space, revealing substantial opportunities for deeper AI integration across scientific disciplines. Furthermore, we explore the potential and challenges of facilitating collaboration between AI and scientific communities through the lens of link prediction. Our findings and tools aim to promote more impactful interdisciplinary collaborations and accelerate scientific discovery through deeper and broader AI integration.

Conversational systems are now capable of producing impressive and generally relevant responses. However, we have no visibility nor control of the socio-emotional strategies behind state-of-the-art Large Language Models (LLMs), which poses a problem in terms of their transparency and thus their trustworthiness for critical applications. Another issue is that current automated metrics are not able to properly evaluate the quality of generated responses beyond the dataset's ground truth. In this paper, we propose a neural architecture that includes an intermediate step in planning socio-emotional strategies before response generation. We compare the performance of open-source baseline LLMs to the outputs of these same models augmented with our planning module. We also contrast the outputs obtained from automated metrics and evaluation results provided by human annotators. We describe a novel evaluation protocol that includes a coarse-grained consistency evaluation, as well as a finer-grained annotation of the responses on various social and emotional criteria. Our study shows that predicting a sequence of expected strategy labels and using this sequence to generate a response yields better results than a direct end-to-end generation scheme. It also highlights the divergences and the limits of current evaluation metrics for generated content. The code for the annotation platform and the annotated data are made publicly available for the evaluation of future models.

Technological advances in artificial intelligence (AI) have enabled the development of large vision language models (LVLMs) that are trained on millions of paired image and text samples. Subsequent research efforts have demonstrated great potential of LVLMs to achieve high performance in medical imaging use cases (e.g., radiology report generation), but there remain barriers that hinder the ability to deploy these solutions broadly. These include the cost of extensive computational requirements for developing large scale models, expertise in the development of sophisticated AI models, and the difficulty in accessing substantially large, high-quality datasets that adequately represent the population in which the LVLM solution is to be deployed. The HOPPR Medical-Grade Platform addresses these barriers by providing powerful computational infrastructure, a suite of foundation models on top of which developers can fine-tune for their specific use cases, and a robust quality management system that sets a standard for evaluating fine-tuned models for deployment in clinical settings. The HOPPR Platform has access to millions of imaging studies and text reports sourced from hundreds of imaging centers from diverse populations to pretrain foundation models and enable use case-specific cohorts for fine-tuning. All data are deidentified and securely stored for HIPAA compliance. Additionally, developers can securely host models on the HOPPR platform and access them via an API to make inferences using these models within established clinical workflows. With the Medical-Grade Platform, HOPPR's mission is to expedite the deployment of LVLM solutions for medical imaging and ultimately optimize radiologist's workflows and meet the growing demands of the field.

As machine learning becomes more widespread and is used in more critical applications, it's important to provide explanations for these models, to prevent unintended behavior. Unfortunately, many current interpretability methods struggle with faithfulness. Therefore, this Ph.D. thesis investigates the question "How to provide and ensure faithful explanations for complex general-purpose neural NLP models?" The main thesis is that we should develop new paradigms in interpretability. This is achieved by first developing solid faithfulness metrics and then applying the lessons learned from this investigation to develop new paradigms. The two new paradigms explored are faithfulness measurable models (FMMs) and self-explanations. The idea in self-explanations is to have large language models explain themselves, we identify that current models are not capable of doing this consistently. However, we suggest how this could be achieved. The idea of FMMs is to create models that are designed such that measuring faithfulness is cheap and precise. This makes it possible to optimize an explanation towards maximum faithfulness, which makes FMMs designed to be explained. We find that FMMs yield explanations that are near theoretical optimal in terms of faithfulness. Overall, from all investigations of faithfulness, results show that post-hoc and intrinsic explanations are by default model and task-dependent. However, this was not the case when using FMMs, even with the same post-hoc explanation methods. This shows, that even simple modifications to the model, such as randomly masking the training dataset, as was done in FMMs, can drastically change the situation and result in consistently faithful explanations. This answers the question of how to provide and ensure faithful explanations.